GCP Snapshot Automation Scripts

Pre- and post-snapshot hooks for Google Cloud Platform (GCP) Compute Engine persistent disk snapshots, ensuring application- and filesystem-consistent backups.

Overview

Taking live disk snapshots without quiescing active databases and filesystems can cause data corruption upon restore. These scripts coordinate a synchronized freeze and thaw across the storage and service stack.

Architecture & Flow

1. gcloud-snapshots/pre.sh (Quiesce):
   • Locks MySQL / MariaDB tables (FLUSH TABLES WITH READ LOCK) in a background session that keeps the lock until post.sh runs, or for at most LOCK_MAX seconds
   • Closes the Elasticsearch index
   • Flushes write buffers (sync) and freezes the XFS filesystem (xfs_freeze -f /)
   • If any step fails, undoes the steps already taken (by running post.sh) and exits non-zero
2. GCP Disk Snapshot: Snapshot captured in a clean, crash-consistent state.
3. gcloud-snapshots/post.sh (Thaw & Resume):
   • Unfreezes XFS filesystem (xfs_freeze -u /)
   • Releases the database lock by ending the session that holds it
   • Reopens the Elasticsearch index
   • Attempts every step even if one fails, and exits non-zero if any failed

Setup

Configure these scripts as pre/post snapshot guest environment scripts in Google Cloud Compute Engine snapshot schedules. Install both in the same directory, since pre.sh runs post.sh to roll back.

Database credentials are read from /root/.my.cnf (a [client] section, mode 600), never from the command line. Without that file the client falls back to its defaults, such as unix-socket authentication for root.

The settings at the top of pre.sh (database client, lock timeouts, Elasticsearch URL and index, mount to freeze, state directory) can be edited there or overridden from the environment. Set MYSQL, ES_INDEX or FREEZE_MOUNT to an empty string to skip that step. STATE_DIR (default /run/gcloud-snapshot) must match in both scripts and must not be on the frozen filesystem.

Guest-flush (application-consistent) snapshots are not supported on Hyperdisk volumes, so these hooks do not run for Hyperdisk; snapshots of those disks are crash-consistent.